AlphaFold 3‑style co‑folding models can predict biomolecular complexes with high structural accuracy, yet many of their outputs violate physical constraints such as chain overlap at interfaces, distorted ligand bond lengths and angles, non‑planar rings, and inverted stereocenters. Existing solutions either steer the sampler with physics‑aware potentials—drastically increasing sampling cost and memory so that inference on large complexes becomes infeasible—or fine‑tune the model, which is time‑consuming and tied to a specific architecture. We observe that physical validity can be fully verified at inference time from quantities already available to the sampler. Consequently we treat it as a constrained inference problem and introduce two closed‑form projection operators applied to the diffusion model’s denoised coordinate estimate $\hat{x}_0$:
- an inter‑chain van der Waals projection that pushes apart the most severely clashing atom pairs;
- a ligand distance‑geometry projection that restores bond lengths, angles, internal contacts, planarity and chirality. Both operators are local, sparse, displacement‑capped, require no network evaluations, gradients or importance sampling, and leave the denoiser and its weights untouched, allowing them to be dropped into any AF3‑style sampler without retraining.
We evaluated the approach on two independently developed models, Boltz‑2 and OpenFold‑3, across five benchmarks (CASP15, CASP16, the PoseBusters monomer and complex sets, and the Boltz physical‑validity test set). The method recovers perfect physical validity while preserving structural‑accuracy and ligand‑placement metrics, with negligible runtime and memory overhead. This provides a practical, model‑agnostic route to physically valid all‑atom structure prediction.
Review